Abstract: lanetary rover systems need to perform terrain segmentation to identify drivable areas and plan the path, so as to ensure the success of rover detection missions. At present, the task of Mars terrain segmentation is difficult and the computational resources of the rover are limited. This paper proposes a lightweight semantic segmentation network based on DeepLab v3+ network structure. The backbone network is MobileNetV2. The Atrous spatial Pyramid pooling (ASPP) module is optimized by dense connection to further expand the receptive field of the atrous convolution. The coordinate attention (CA) mechanism proposed recently is used to increase the feature extraction ability of our network. CA DeepLab v3+ network is verified by AI4Mars public dataset, which shows that the recall rate of the algorithm can reach 91%, 92%, 89% and 75% in soil, bedrock, sand and large rock, respectively.

Key words: Mars terrain image, DeepLab v3+ network, coordinate attention, semantic segmentation, rover